Method for predicting trajectory of aerial target

Abstract

To solve the problem that the existing air dynamic target trajectory prediction method is difficult to meet the requirements of the autonomous air refueling terminal docking in multiple dimensions such as prediction accuracy, accuracy maintenance ability in long prediction time domain, calculation real-time, robustness and generalization ability in complex disturbance environment, the present application proposes an air target trajectory prediction method, which uses a historical sliding window with a length of N+1 frames and a step of 1 frame to slide sample the target measured three-dimensional absolute coordinates; 6<=N<=30; the adjacent frame difference of the N+1 frame three-dimensional absolute coordinate time sequence sampled at present is processed, N frame three-dimensional displacement increment time sequence is obtained and normalized, and then input into the pre-trained trajectory prediction model, the model outputs the normalized three-dimensional displacement increment prediction value of the next frame; the normalized three-dimensional displacement increment prediction value of the next frame is de-normalized, and added to the latest one frame three-dimensional absolute coordinate sampled at present, to obtain the three-dimensional absolute coordinate prediction value of the target in the next frame.

Description

Technical Field

[0001] This invention relates to a method for predicting the trajectory of aerial targets. Background Technology

[0002] During the terminal docking phase of autonomous aerial refueling using a hose-and-drogue system, the relative motion between the receiver aircraft and the refueling drogue exhibits highly dynamic, nonlinear, and strongly coupled complex characteristics. The receiver aircraft, flying at high speed, needs to precisely track and dock with the drogue towed by the tanker. However, the drogue itself is affected by multiple disturbances, including atmospheric turbulence, the tanker's wake, and the elastic oscillations of the hose, resulting in a significantly random and unpredictable trajectory. This complex relative motion presents a severe challenge to traditional single-frame pose measurement methods: measurement noise, processing delays, and sensor dynamic response lags often lead to time delays in the real-time acquired pose information. Directly using the measurement results of the current frame for control decisions is insufficient to meet the real-time requirements of high-precision docking, easily resulting in inaccurate measurements, failure to keep up, and misalignment.

[0003] Currently, relative navigation technologies applied to autonomous aerial refueling mainly rely on the following types of schemes, but they all have inherent limitations that are difficult to overcome in handling dynamic target trajectory prediction:

[0004] Filtering-based estimation methods, such as Kalman filtering (KF), extended Kalman filtering (EKF), and unscented Kalman filtering (UKF), recursively filter historical measurement data by establishing a target motion model (such as a constant velocity model or a constant acceleration model) to estimate the current state and predict the future state. However, their predictive performance heavily depends on the accuracy of the preset motion model. When the cone sleeve undergoes nonlinear and non-Gaussian motion due to complex airflow disturbances, a motion model with fixed parameters struggles to accurately describe its true dynamic characteristics, leading to a rapid accumulation of prediction errors over time. The accuracy drops sharply over longer prediction time domains, failing to meet the stringent accuracy requirements of the final docking stage.

[0005] Traditional time series analysis methods, such as autoregressive (AR) models, moving average (MA) models, and their combinations (ARMA, ARIMA), attempt to extract statistical patterns from historical pose sequences for extrapolation and prediction. However, conical motion is affected by multi-source random disturbances, exhibiting non-stationary and nonlinear time-varying characteristics. Traditional time series models assume that the data satisfies stationarity or can be stationarized through differencing, making it difficult to capture the complex dynamic patterns and long-term dependencies in conical motion, resulting in poor prediction robustness under strong disturbance conditions.

[0006] Prediction methods based on shallow machine learning, such as support vector regression (SVR) and shallow neural networks, while possessing some nonlinear fitting capabilities, are limited by model capacity and feature extraction capabilities, making it difficult to effectively mine deep spatiotemporal features and high-order motion patterns in continuous frame pose data. Furthermore, these methods typically require tedious manual feature engineering, have limited generalization ability, and struggle to adapt to the complex and variable aerodynamic environment and lighting conditions during aerial refueling.

[0007] Single-frame deep learning-based pose estimation methods: In recent years, single-frame pose estimation methods based on convolutional neural networks (CNNs) have made significant progress in accuracy. However, they are essentially still mappings of the static image at the current moment, lacking explicit modeling of temporal information. Directly extending these methods to multi-frame prediction often employs simple post-processing or moving average strategies, failing to fully utilize the temporal correlation of historical trajectories and effectively learning the dynamic evolution of conical motion. Consequently, they lack foresight and accuracy in predicting poses in future frames.

[0008] In summary, existing methods for predicting the trajectory of dynamic targets in the air are insufficient to simultaneously meet the stringent requirements of autonomous aerial refueling terminal docking in terms of prediction accuracy, accuracy maintenance over a long prediction time domain, real-time computation, and robustness and generalization ability under complex disturbance environments. Summary of the Invention

[0009] To address the technical challenge of existing aerial dynamic target trajectory prediction methods simultaneously meeting the requirements for autonomous aerial refueling terminal docking across multiple dimensions, including prediction accuracy, accuracy maintenance over long prediction time domains, computational real-time performance, and robustness and generalization capability under complex disturbance environments, this invention proposes an aerial target trajectory prediction method.

[0010] The technical solution of this invention is:

[0011] The method for predicting the trajectory of aerial targets is unique in that it includes the following steps:

[0012] The measured three-dimensional absolute coordinates of the target are sampled using a historical sliding window with a length of N+1 frames and a step size of 1 frame; 6≤N≤30;

[0013] The N+1 frames of the three-dimensional absolute coordinate time sequence sampled at the current time are differentially processed to obtain the N frames of the three-dimensional displacement increment time sequence at the current time.

[0014] After normalizing the N-frame three-dimensional displacement increment time sequence, it is input into the pre-trained trajectory prediction model, and the trajectory prediction model outputs the normalized three-dimensional displacement increment prediction value of the next frame.

[0015] After inverse normalization of the predicted normalized 3D displacement increment for the next frame, it is added to the latest 3D absolute coordinates sampled at the current time to obtain the predicted 3D absolute coordinates of the target in the next frame.

[0016] Furthermore, the trajectory prediction model includes a temporal encoder, a multi-head temporal attention module, a linear projection layer, a feature fusion module, and a fully connected block;

[0017] The timing encoder uses a 4-layer LSTM network to perform deep timing encoding on the input normalized displacement increment timing sequence in order to capture the sequential dependencies of the displacement increment data on the time axis and finally output the hidden state sequence.

[0018] The multi-head temporal attention module is used to extract the most valuable unified global contextual features for the prediction target from the hidden state sequence;

[0019] The linear projection layer is used to linearly map the latest frame displacement increment in the normalized displacement increment time sequence and output a linear projection feature vector.

[0020] The feature fusion module is used to add the projected feature vector output by the linear projection layer to the unified global context feature output by the multi-head temporal attention module element by element, and output the fused feature.

[0021] Fully connected blocks are used to map fused features to normalized displacement increments for the next frame.

[0022] Furthermore, the multi-head temporal attention module includes a parallel attention computation unit, a residual connection unit, and a layer normalization unit;

[0023] The parallel attention computation unit uses m parallel attention heads to perform parallel computation and weighted aggregation of the importance of each historical frame relative to the prediction target within the sliding historical window, and obtains the reconstructed features;

[0024] The residual connection unit is used to add the reconstructed features output by the parallel attention computing unit to the features of the corresponding time step in the hidden state sequence element by element, and output the residual fusion features.

[0025] The layer normalization unit is used to normalize the residual fusion features output by the residual connection unit, and then take the context features of the last time step as the unified global context features and output them.

[0026] Furthermore, the method for constructing the training and test sets for training and testing the trajectory prediction model is as follows:

[0027] Conduct target pose localization experiments to obtain a continuous three-dimensional absolute coordinate time sequence of the target. ; ;

[0028] Calculate continuous three-dimensional absolute coordinate time series The target displacement increment between adjacent frames is used to obtain a three-dimensional displacement increment time sequence. ;

[0029] Three-dimensional displacement increment time series Normalization is performed to obtain the normalized displacement increment time series. ;

[0030] Set a sliding window with a length of N frames and a step size of 1 frame for the normalized displacement increment time series. Sliding sampling to obtain sample data , , ..., ;

[0031] Extracting sample data 3D displacement increment of the next frame As with Corresponding tags Extract sample data 3D displacement increment of the next frame As with Corresponding tags Extract sample data 3D displacement increment of the next frame As with Corresponding tags Extract sampling data; ... 3D displacement increment of the next frame As with Corresponding tags ;

[0032] Construct a dataset using sample data and its corresponding labels;

[0033] The dataset is divided into training and testing sets according to a preset ratio.

[0034] Furthermore, for the three-dimensional displacement increment time series... The method for normalization is:

[0035] Statistical three-dimensional displacement increment time series The maximum and minimum values ​​of the displacement increments along the x, y, and z axes are obtained, and the maximum and minimum values ​​of the x, y, and z axes are used to construct the MinMax normalizers for the x-axis, y-axis, and z-axis, respectively.

[0036] Three-dimensional displacement increment time series The inputs are fed into the X-axis MinMax normalizer, Y-axis MinMax normalizer, and Z-axis MinMax normalizer to obtain the normalized results for the X, Y, and Z axes. The normalized results for the X, Y, and Z axes are then concatenated to obtain the normalized displacement increment time series. .

[0037] Furthermore, the training method for the trajectory prediction model is as follows:

[0038] The trajectory prediction model is trained using a pre-established training set, supervised by the Huber loss function during training, and updated with the network parameters of the trajectory prediction model through backpropagation algorithm, AdamW optimizer and gradient pruning strategy.

[0039] Furthermore, Huber's loss function employs a piecewise strategy to calculate the loss value:

[0040] When the absolute value of the prediction error of the trajectory prediction model is less than or equal to the preset threshold δ, the Huber loss is calculated as a quadratic term.

[0041] When the absolute value of the prediction error of the trajectory prediction model is greater than the preset threshold δ, the Huber loss is calculated as a linear term.

[0042] This invention also proposes an aerial target trajectory prediction system, which is characterized by including a data sampling unit, a data preprocessing unit, a trajectory prediction model, and a coordinate restoration unit.

[0043] The data sampling unit is used for sliding sampling of the measured three-dimensional absolute coordinates of the target, and at each sampling time, N+1 frames of three-dimensional absolute coordinate time sequence are obtained; 6≤N≤30;

[0044] The data preprocessing unit includes a difference calculation module and a normalization processing module. The difference calculation module is used to perform difference calculation on adjacent frames in the N+1 frames of the three-dimensional absolute coordinate time sequence at each sampling time to obtain the N frames of the three-dimensional displacement increment time sequence at each sampling time. The normalization processing module is used to normalize the N frames of the three-dimensional displacement increment time sequence at each sampling time to obtain the normalized three-dimensional displacement increment time sequence.

[0045] The trajectory prediction model adopts the trajectory prediction model used in the above-mentioned aerial target trajectory prediction method, and is used to predict the normalized three-dimensional displacement increment of the next frame based on the normalized three-dimensional displacement increment time sequence of the current moment.

[0046] The coordinate restoration unit is used to inversely normalize the predicted normalized 3D displacement increment for the next frame, add it to the latest 3D absolute coordinates of the current frame, and output the predicted 3D absolute coordinates of the target for the next frame.

[0047] Furthermore, the normalization module consists of an X-axis MinMax normalizer, a Y-axis MinMax normalizer, and a Z-axis MinMax normalizer, and its construction method is as follows:

[0048] Conduct target pose localization experiments to obtain a continuous three-dimensional absolute coordinate time sequence of the target. ; ;

[0049] Calculate continuous three-dimensional absolute coordinate time series The target displacement increment between adjacent frames is used to obtain a three-dimensional displacement increment time sequence. ;

[0050] Statistical three-dimensional displacement increment time series The maximum and minimum values ​​of the displacement increments along the x, y, and z axes are obtained, and the maximum and minimum values ​​of the x, y, and z axes are used to construct the MinMax normalizers for the x-axis, y-axis, and z-axis, respectively.

[0051] This invention also proposes a trajectory prediction model, which is special in that it adopts the trajectory prediction model used in the above-mentioned aerial target trajectory prediction method.

[0052] Compared with the prior art, the present invention has the following significant advantages:

[0053] 1. High prediction accuracy

[0054] To address the challenges of "long-distance forgetting" and limited feature extraction in traditional time-series models when processing complex flight trajectories, this invention combines LSTM deep sequences with a multi-head temporal attention mechanism to establish a trajectory prediction model. By analyzing historical displacement increment sequences in parallel using multiple independent attention heads (preferably 8 groups), the trajectory prediction model can simultaneously focus on local minor perturbations and global motion trends, uncovering complex nonlinear features hidden in historical trajectories. This accurately identifies the potential motion patterns and trends of the target under complex conditions such as acceleration changes and turns, significantly improving the prediction accuracy of the 3D position increment at the next sampling moment, thereby enhancing the prediction accuracy of the 3D position at the next sampling moment and providing crucial lead time for the flight control system.

[0055] 2. High predictive reliability under long-term operation

[0056] This invention employs a physical anchoring and restoration logic of "next frame displacement increment + latest observed coordinates" to ensure that the predicted three-dimensional absolute coordinates of each frame are based on the latest actual observation values ​​as the physical reference. This suppresses the trajectory drift problem that commonly occurs over time in pure sequence generation models and ensures the reliability of the continuous rolling prediction process.

[0057] 3. Good noise resistance and operational stability

[0058] This invention utilizes the temporal modeling capabilities of LSTM networks to filter high-frequency noise, effectively suppressing input coordinate "jump" noise interference caused by factors such as sudden changes in illumination and occlusion during upstream visual or sensor measurements. Simultaneously, the Huber loss function reduces the impact of outliers on the gradient of the trajectory prediction model. Through the synergy of the LSTM network and the Huber loss function, the trajectory prediction model maintains robustness even with fluctuations in upstream sensor data quality. Finally, physical anchoring restores a smooth trajectory that conforms to the laws of physical motion, significantly improving the usability of the prediction results. Experimental data shows that even with instantaneous noise in the input 3D position increment, the trajectory output by the method of this invention remains highly smooth, avoiding instability caused by severe fluctuations in the input signal.

[0059] 4. Good generalization ability

[0060] This invention employs a Delta modeling logic, calculating and predicting the 3D spatial displacement increment between adjacent frames instead of directly predicting absolute coordinates. This allows the model to learn the target's motion patterns rather than its specific spatial location. This gives the model strong generalization ability; regardless of the target's initial coordinates, the model can output consistent and reliable displacement increment predictions for the next frame. Therefore, for prediction tasks with targets at different initial positions and distance distributions, there is no need for repeated training in specific spatial domains, improving the model's deployment flexibility.

[0061] 5. Good real-time response capability

[0062] This invention optimizes the model parameter scale and computation graph structure to ensure low inference latency on the airborne computing platform. By designing N-frame historical displacement increment inputs and single-step displacement increment outputs, it provides response lead-in to the flight control unit, effectively compensating for system latency in the vision link and providing high-frequency predictive commands for precise and stable autonomous docking control. According to measured data, the average processing time of this invention is approximately 1.01ms, with an equivalent frame rate exceeding 980Hz. This high-frequency output capability not only reduces computational latency in the prediction stage but also provides computational margin for complex flight control closed-loop control.

[0063] 6. The method of the present invention is not limited to predicting the trajectory of refueling cones, but is also applicable to the three-dimensional trajectory prediction task of other aerial moving targets such as UAVs, target drones, and birds. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0065] Figure 2This is a schematic diagram of the trajectory prediction model structure.

[0066] Figure 3 The image shows the prediction results of the trajectory prediction model. Detailed Implementation

[0067] The present invention will be further described below with reference to the accompanying drawings.

[0068] Reference Figure 1 The aerial target trajectory prediction method proposed in this invention includes the following steps:

[0069] Step 1: Collect and process data to build training and test sets;

[0070] Step 1.1, collect data;

[0071] A target pose localization experiment was conducted, during which a motion capture system was used to acquire a continuous three-dimensional absolute coordinate time sequence of the target. ,in The three-dimensional absolute coordinates of the i-th frame;

[0072] Step 1.2, Differential processing;

[0073] Calculating continuous three-dimensional absolute coordinate time series using difference operators Target displacement increment between adjacent frames The three-dimensional displacement increment time series was obtained by arranging the data in chronological order. .

[0074] Through the above differential processing, the prediction object of the subsequent trajectory prediction model is transformed from the absolute coordinate of the next frame to the displacement increment of the absolute coordinate of the next frame relative to the absolute coordinate of the current frame. This removes the absolute position information of the target, allowing the subsequent model training to focus on learning the essential motion laws of the target, such as inertia, acceleration changes, and local disturbances, thereby enhancing the model's generalization ability in different spatial regions.

[0075] Step 1.3, MinMax normalizer construction;

[0076] Constructing a MinMax normalizer: statistical analysis of three-dimensional displacement increment time series The maximum and minimum values ​​of the displacement increments along the x, y, and z axes are obtained, and the statistical maximum and minimum values ​​of these values ​​are used to construct the MinMax normalizers for the x, y, and z axes, respectively. Each axis's MinMax normalizer maps the displacement increments of each coordinate axis to the [-1, 1] interval to eliminate the impact of numerical scale differences between different coordinate axes on subsequent model training. It also saves the maximum and minimum values ​​of the displacement increments for each coordinate axis, the range of the displacement increments (i.e., the numerical variation range of the displacement increments), and the scaling factor for each coordinate axis to support model training, inference, and denormalization.

[0077] Save the constructed MinMax normalizers for each axis for reuse in subsequent online inference.

[0078] Step 1.4, data normalization processing;

[0079] The three-dimensional displacement increment time series obtained in step 1.2 Input the MinMax normalizers constructed in step 1.3 for each axis. The MinMax normalizers for each axis output the normalized results of the data for each axis. Concatenate the normalized results of each axis to obtain the normalized displacement increment time series. .

[0080] Step 1.5, construct the dataset;

[0081] Step 1.5.1, Set the sampling window: Set a sliding window with a length of N frames and a step size of 1 frame, where 6 ≤ N ≤ 30. In this embodiment, N=20.

[0082] Step 1.5.2: Construct supervised samples and labels based on the normalized displacement increment time series:

[0083] Using the set sampling window, the normalized displacement increment time series Perform sliding sampling to obtain sample data , , ..., ;

[0084] Extracting sample data 3D displacement increment of the next frame As with Corresponding tags Extract sample data 3D displacement increment of the next frame As with Corresponding tags Extract sample data 3D displacement increment of the next frame As with Corresponding tags Extract sampling data; ... 3D displacement increment of the next frame As with Corresponding tags ;

[0085] Step 1.5.3: Construct a dataset using the sample data and its corresponding labels.

[0086] Step 1.6, Split the dataset;

[0087] The dataset is divided into training and testing sets according to a preset ratio.

[0088] Step 2: Establish a trajectory prediction model;

[0089] The input to the trajectory prediction model is a normalized displacement increment sequence. This normalized displacement increment time series sequence is input into the model in the form of a tensor with shape (batch, N, 3), where N represents the number of historical displacement increment frames and 3 represents the three coordinate axes x, y, and z. In this embodiment, N=20, and the normalized displacement increment sequence is input into the model in the form of a tensor with shape (batch, 20, 3).

[0090] Reference Figure 2 The trajectory prediction model includes a temporal encoder, a multi-head temporal attention module, a linear projection layer, a feature fusion module, and a fully connected block.

[0091] The timing encoder employs a 4-layer LSTM network to perform deep timing encoding on the input normalized displacement increment timing sequence, in order to capture the sequential dependencies of the displacement increment data on the time axis. Finally, the 4-layer LSTM network outputs a hidden state sequence with a shape of (batch, N, 128), which serves as the input to the multi-head temporal attention module. In this embodiment, the 4-layer LSTM network outputs a hidden state sequence with a shape of (batch, 20, 128).

[0092] The multi-head temporal attention module is used to extract the most valuable unified global contextual features for the prediction target from the hidden state sequence.

[0093] The multi-head temporal attention module includes a parallel attention computation unit, a residual connection unit, and a layer normalization unit.

[0094] The parallel attention calculation unit uses m parallel attention heads (preferably 8 in this embodiment) to perform parallel calculation and weighted aggregation of the importance of each historical frame relative to the prediction target within the sliding historical window, and obtains the reconstructed features;

[0095] The residual connection unit is used to add the reconstructed features output by the parallel attention computation unit to the features of the corresponding time step in the hidden state sequence (original input) element by element, and output residual fusion features to preserve the original temporal information and alleviate the gradient vanishing problem in deep networks.

[0096] The layer normalization unit is used to normalize the residual fusion features output by the residual connection unit, and then take the context features of the last time step as the unified global context features and output them.

[0097] The linear projection layer is used to linearly map the latest frame displacement increment in the normalized displacement increment time sequence to obtain a linear projection feature vector of shape (batch, 128) and output it.

[0098] The feature fusion module is used to add the projected feature vector output by the linear projection layer to the unified global context feature output by the multi-head temporal attention module element by element, and output a fused feature with shape (batch, 128).

[0099] The fully connected block is used to map the fused features to the normalized displacement increment for the next frame. In this embodiment, the fully connected block includes a first fully connected sub-layer, an activation sub-layer, a random deactivation sub-layer, and a second fully connected sub-layer connected in sequence. The first fully connected sub-layer is used to linearly map the fused features from 128 dimensions to 64-dimensional intermediate features. The activation sub-layer is used to perform ReLU non-linear activation processing on the 64-dimensional intermediate features and output the activated features. The random deactivation sub-layer is used to randomly discard the activated features with a deactivation probability of 0.2 and output the features after random deactivation. The second fully connected sub-layer is used to linearly map the features after random deactivation from 64 dimensions to 3 dimensions and output the normalized displacement increment for the next frame.

[0100] It should be noted that the 128-dimensional hidden state sequence is only an optional feature of this invention and is not mandatory. In other embodiments, those skilled in the art can adjust the dimension of the hidden state sequence according to the trajectory sequence length, prediction accuracy requirements, and computational resource constraints; correspondingly, the dimensions of the unified global context features, fusion features, and projected feature vectors generated based on the hidden state sequence will be adaptively adjusted as the dimension of the hidden state sequence changes.

[0101] Step 3: Train the trajectory prediction model;

[0102] Step 3.1, Model Training;

[0103] The trajectory prediction model is trained using a training set, with Huber loss function used for supervision during training. The network parameters of the trajectory prediction model are updated through backpropagation algorithm, AdamW optimizer and gradient pruning strategy, so that the trajectory prediction model can have high prediction accuracy and resist occasional jump noise generated by upstream visual measurement or sensor measurement.

[0104] The Huber loss function employs a piecewise strategy to calculate the loss value:

[0105] When the absolute value of the prediction error of the trajectory prediction model is less than or equal to the preset threshold δ, the Huber loss is calculated as a quadratic term. This calculation method approximates the mean square error (MSE), which is beneficial to improving the prediction accuracy of the trajectory prediction model in the small error range.

[0106] When the absolute value of the prediction error of the trajectory prediction model is greater than the preset threshold δ, the Huber loss is calculated as a linear term. This calculation method approximates the mean absolute error (MAE), which helps to reduce the interference of outliers on the model weight update.

[0107] Step 3.2, Model performance evaluation;

[0108] The trained trajectory prediction model is tested using a test set. Based on the test results, the average Euclidean error, maximum Euclidean error, minimum Euclidean error, stability index, average inference time, and equivalent frame rate are used as evaluation indicators to assess the prediction performance of the trajectory prediction model. If the prediction performance does not meet the requirements, return to step 3.1 to retrain until the prediction performance of the trajectory prediction model meets the requirements.

[0109] Step 4: Trajectory prediction and physical anchoring reconstruction;

[0110] Step 4.1, displacement increment reasoning;

[0111] In actual prediction, a historical sliding window with a length of N+1 frames and a step size of 1 frame is set. The three-dimensional absolute coordinates of the target, acquired using the existing airborne vision measurement system, are sampled. At the current moment, the latest N+1 frames (21 frames in this embodiment, N=20) of the three-dimensional absolute coordinate time series are acquired. The same method as in step 1.2 is used to perform differential processing on the latest N+1 frames (21 frames in this embodiment) of the three-dimensional absolute coordinate time series, resulting in an N-frame (20 frames in this embodiment) displacement increment time series. The MinMax normalizer for each axis constructed in step 1.3 is called, and the N-frame (20 frames in this embodiment) displacement increment time series is input into it to obtain an N-frame (20 frames in this embodiment) normalized displacement increment sequence. The N-frame (20 frames in this embodiment) normalized displacement increment sequence is input into the trained and evaluated trajectory prediction model. The trajectory prediction model outputs the predicted normalized displacement increment value (a dimensionless normalized value) for the next frame at the current sampling moment, which can be denoted as... .

[0112] Step 4.2, coordinate restoration;

[0113] The normalized displacement increment of the trajectory prediction model output for the next frame is obtained through inverse normalization. The displacement increment of the next frame is restored to physical units, with the unit being mm / frame. The displacement increment of the next frame in physical units is added to the three-dimensional absolute coordinates of the latest frame currently sampled to obtain the predicted value of the three-dimensional absolute coordinates of the target at the next sampling time (i.e., the t+1th frame) (in mm). The three-dimensional absolute coordinates can provide a reference for the delay compensation and control decision of the flight control system.

[0114] Step 5: The historical sliding window is updated in real time to achieve trajectory prediction;

[0115] For each new measured 3D absolute coordinate of the target in a frame, the oldest 3D absolute coordinate in the historical sliding window is removed and the latest 3D absolute coordinate is added to the historical sliding window. The 3D absolute coordinate of the target in the latest N+1 frames (21 frames in this embodiment) is reconstructed. The steps of difference, normalization, model prediction and coordinate restoration are executed in a loop to achieve real-time rolling trajectory prediction.

[0116] Example:

[0117] Hardware configuration:

[0118] Computing Platform: This embodiment uses an NVIDIA Jetson AGX Orin (32GB) embedded high-performance computing module as the core for algorithm execution. This platform has powerful Tensor Core computing power, which can efficiently support real-time inference of deep neural networks.

[0119] Input source: The data source is a continuous three-dimensional absolute coordinate sequence of the target output by the receiving airborne vision measurement system. The sampling frequency is set to 100Hz to match the control requirements during the high-speed docking process.

[0120] Software implementation:

[0121] Trajectory prediction model structure:

[0122] Temporal coding layer: A 4-layer LSTM network is used, with the hidden layer dimension of the LSTM network set to 128. The temporal coding layer is responsible for extracting deep temporal dependency features from 20 consecutive frames of historical displacement increment data and outputting the hidden state sequence.

[0123] Multi-head temporal attention module: Employs a multi-head attention mechanism, with eight attention heads in this embodiment. By computing the attention weights of different heads in parallel, it simultaneously captures local micro-oscillations and global motion trends to generate unified global contextual features.

[0124] Linear projection layer: linearly maps the latest frame displacement increment in the normalized displacement increment sequence, and outputs a linear projection feature vector with shape (batch, 128).

[0125] Feature fusion module: Introducing residual connection, which sums the linear projection feature vector of the displacement increment of the latest frame with the attention aggregation feature element by element, enhancing the model’s sensitivity to the current motion state.

[0126] The fully connected block consists of two fully connected sub-layers (128→64→3), a ReLU activation function, and a random deactivation sub-layer, and finally outputs the predicted normalized shift increment for the next frame.

[0127] The loss function used is Huber Loss (δ=1.0), which is highly robust when dealing with jump noise generated by visual or sensor measurements.

[0128] The optimization algorithm uses the AdamW optimizer with a learning rate of 0.0005 and incorporates weight decay (weight_decay=1e-3) to improve the stability of the trajectory prediction model.

[0129] Training data construction:

[0130] The continuous three-dimensional absolute coordinates in each training trajectory are sorted by time and then differentially divided to obtain a three-dimensional displacement increment sequence. After normalizing the displacement increment sequence, samples and labels are constructed using a sliding window method, that is, every 20 consecutive frames of displacement increment constitutes a sample, and the displacement increment of the next frame immediately following it is used as the label of the sample. All samples are divided into training set and test set according to the ratio of 80% and 20%.

[0131] Online reasoning process:

[0132] After the qualified trajectory prediction model is actually deployed, the latest 21 frames of the target's three-dimensional absolute coordinates are received each time. The displacement increment of 20 frames is first obtained by difference, then normalized and input into the trajectory prediction model. The trajectory prediction model outputs the normalized displacement increment of the next frame. After inverse normalization, the displacement increment of the next frame in mm / frame is obtained. The displacement increment is added to the current latest frame of three-dimensional absolute coordinates to obtain the three-dimensional absolute coordinate prediction result of the next sampling time.

[0133] Experimental verification:

[0134] During the testing process, the average Euclidean error, maximum Euclidean error, minimum Euclidean error, stability index, average inference time, and equivalent frame rate were used as evaluation metrics. Figure 3 The test results shown in the figure are compared. The Ground Truth is the data collected by the motion capture system in the target pose localization experiment. It can be concluded that the average prediction error of the present invention is reduced by about 75% to 80% compared with the EKF algorithm, the average processing time is about 1.01ms, and the equivalent frame rate reaches more than 980Hz.

Claims

1. A method for predicting the trajectory of aerial targets, characterized in that, Including the following steps: The measured three-dimensional absolute coordinates of the target are sampled using a historical sliding window with a length of N+1 frames and a step size of 1 frame; 6≤N≤30; The N+1 frames of the three-dimensional absolute coordinate time sequence sampled at the current time are differentially processed to obtain the N frames of the three-dimensional displacement increment time sequence at the current time. After normalizing the N-frame three-dimensional displacement increment time sequence, it is input into the pre-trained trajectory prediction model, and the trajectory prediction model outputs the normalized three-dimensional displacement increment prediction value of the next frame. After inverse normalization of the predicted normalized 3D displacement increment for the next frame, it is added to the latest 3D absolute coordinates sampled at the current time to obtain the predicted 3D absolute coordinates of the target in the next frame.

2. The aerial target trajectory prediction method according to claim 1, characterized in that, The trajectory prediction model includes a temporal encoder, a multi-head temporal attention module, a linear projection layer, a feature fusion module, and a fully connected block; The timing encoder uses a 4-layer LSTM network to perform deep timing encoding on the input normalized displacement increment timing sequence in order to capture the sequential dependencies of the displacement increment data on the time axis and finally output the hidden state sequence. The multi-head temporal attention module is used to extract the most valuable unified global contextual features for the prediction target from the hidden state sequence; The linear projection layer is used to linearly map the latest frame displacement increment in the normalized displacement increment time sequence and output a linear projection feature vector. The feature fusion module is used to add the projected feature vector output by the linear projection layer to the unified global context feature output by the multi-head temporal attention module element by element, and output the fused feature. Fully connected blocks are used to map fused features to normalized displacement increments for the next frame.

3. The aerial target trajectory prediction method according to claim 2, characterized in that, The multi-head temporal attention module includes a parallel attention computation unit, a residual connection unit, and a layer normalization unit; The parallel attention computation unit uses m parallel attention heads to perform parallel computation and weighted aggregation of the importance of each historical frame relative to the prediction target within the sliding historical window, and obtains the reconstructed features; The residual connection unit is used to add the reconstructed features output by the parallel attention computing unit to the features of the corresponding time step in the hidden state sequence element by element, and output the residual fusion features. The layer normalization unit is used to normalize the residual fusion features output by the residual connection unit, and then take the context features of the last time step as the unified global context features and output them.

4. The aerial target trajectory prediction method according to claim 1, 2, or 3, characterized in that, The method for constructing the training and test sets for training and testing the trajectory prediction model is as follows: Conduct target pose localization experiments to obtain a continuous three-dimensional absolute coordinate time sequence of the target. ; ; Calculate continuous three-dimensional absolute coordinate time series The target displacement increment between adjacent frames is used to obtain a three-dimensional displacement increment time sequence. ; Three-dimensional displacement increment time series Normalization is performed to obtain the normalized displacement increment time series. ; Set a sliding window with a length of N frames and a step size of 1 frame for the normalized displacement increment time series. Sliding sampling to obtain sample data , , ..., ; Extracting sample data 3D displacement increment of the next frame As with Corresponding tags Extract sample data 3D displacement increment of the next frame As with Corresponding tags Extract sample data 3D displacement increment of the next frame As with Corresponding tags Extract sampling data; ... 3D displacement increment of the next frame As with Corresponding tags ; Construct a dataset using sample data and its corresponding labels; The dataset is divided into training and testing sets according to a preset ratio.

5. The aerial target trajectory prediction method according to claim 4, characterized in that, Three-dimensional displacement increment time series The method for normalization is: Statistical three-dimensional displacement increment time series The maximum and minimum values ​​of the displacement increments along the x, y, and z axes are obtained, and the maximum and minimum values ​​of the x, y, and z axes are used to construct the MinMax normalizers for the x-axis, y-axis, and z-axis, respectively. Three-dimensional displacement increment time series The inputs are fed into the X-axis MinMax normalizer, Y-axis MinMax normalizer, and Z-axis MinMax normalizer to obtain the normalized results for the X, Y, and Z axes. The normalized results for the X, Y, and Z axes are then concatenated to obtain the normalized displacement increment time series. .

6. The aerial target trajectory prediction method according to claim 5, characterized in that, The training method for the trajectory prediction model is as follows: The trajectory prediction model is trained using a pre-established training set, supervised by the Huber loss function during training, and updated with the network parameters of the trajectory prediction model through backpropagation algorithm, AdamW optimizer and gradient pruning strategy.

7. The aerial target trajectory prediction method according to claim 6, characterized in that, The Huber loss function uses a piecewise strategy to calculate the loss value: When the absolute value of the prediction error of the trajectory prediction model is less than or equal to the preset threshold δ, the Huber loss is calculated as a quadratic term. When the absolute value of the prediction error of the trajectory prediction model is greater than the preset threshold δ, the Huber loss is calculated as a linear term.

8. An aerial target trajectory prediction system, characterized in that, It includes a data sampling unit, a data preprocessing unit, a trajectory prediction model, and a coordinate restoration unit; The data sampling unit is used for sliding sampling of the measured three-dimensional absolute coordinates of the target, and at each sampling time, N+1 frames of three-dimensional absolute coordinate time sequence are obtained; 6≤N≤30; The data preprocessing unit includes a difference calculation module and a normalization processing module. The difference calculation module is used to perform difference calculation on adjacent frames in the N+1 frames of the three-dimensional absolute coordinate time sequence at each sampling time to obtain the N frames of the three-dimensional displacement increment time sequence at each sampling time. The normalization processing module is used to normalize the N frames of the three-dimensional displacement increment time sequence at each sampling time to obtain the normalized three-dimensional displacement increment time sequence. The trajectory prediction model adopts the trajectory prediction model used in the aerial target trajectory prediction method according to any one of claims 1-7, and is used to predict the normalized three-dimensional displacement increment of the next frame based on the normalized three-dimensional displacement increment time sequence of the current moment. The coordinate restoration unit is used to inversely normalize the predicted normalized 3D displacement increment for the next frame, add it to the latest 3D absolute coordinates of the current frame, and output the predicted 3D absolute coordinates of the target for the next frame.

9. The aerial target trajectory prediction system according to claim 8, characterized in that, The normalization module consists of an X-axis MinMax normalizer, a Y-axis MinMax normalizer, and a Z-axis MinMax normalizer. Its construction method is as follows: Conduct target pose localization experiments to obtain a continuous three-dimensional absolute coordinate time sequence of the target. ; ; Calculate continuous three-dimensional absolute coordinate time series The target displacement increment between adjacent frames is used to obtain a three-dimensional displacement increment time sequence. ; Statistical three-dimensional displacement increment time series The maximum and minimum values ​​of the displacement increments along the x, y, and z axes are obtained, and the maximum and minimum values ​​of the x, y, and z axes are used to construct the MinMax normalizers for the x-axis, y-axis, and z-axis, respectively.

10. A trajectory prediction model, characterized in that, The trajectory prediction model used in the aerial target trajectory prediction method according to any one of claims 1-7 is adopted.

Images